[Mesa-dev] [RFC] intel: add support for EGL_KHR_gl_image
eric at anholt.net
Thu Jan 3 17:00:39 PST 2013
Abdiel Janulgue <abdiel.janulgue at linux.intel.com> writes:
> On Friday, December 14, 2012 11:38:04 AM Eric Anholt wrote:
>> Abdiel Janulgue <abdiel.janulgue at linux.intel.com> writes:
>> > This patch set adds support for KHR_gl_texture_2D_image,
>> > KHR_gl_texture_cubemap_image and KHR_gl_texture_3D_image for Gen 4 ->
>> > Gen7 HW. The extension enables us to be a bit more conformant to the
>> > spec in Android at least. I've tested it on SandyBridge and IvyBridge.
>> > The approach is to export a mipmap level to a DRIImage accounting for
>> > the level's offset from the surface base address including the x and y
>> > offsets on cubemap face or 3D texture slices.
>> > I've pushed it in git at gitorious.org/mesa3d/mesa.git KHR_gl_image
>> > branch. I'll submit a a separate piglit test to soon to piglit-dev.
>> Cool. We really should have these extensions.
>> But we also really need some piglit tests to stress this.
> Thanks! I just submitted an initial test to piglit list.
>> intel_setup_image_from_mipmap_tree() code in this series can orphan
>> siblings the moment they're created. intel_finalize_mipmap_tree() can
>> also do so currently. We need to be able to track somehow where the
>> canonical miptree contents are and resolve our work back to it.
> Good catch. Do you think if it would make sense to drop support for this non-
> tile-aligned <= gen4 hack in this case? It would make the code much more
> simpler. The older hardware doesn't seem to support rendering to the fine
> grained X and Y tile offsets anyways - which is needed to resolve back to the
> specific miptree level.
Sorry for the delay, I think we all disappeared over the holidays.
Let's see if I've got this all right. It looks like today we have
* turn an EGLImage into a 0-level texture
* turn an EGLImage into a renderbuffer
* Create an EGLImage from a format/width/height/drm usage
* possibly create an EGLImage from a format/bo name/stride
* turn an EGLImage into a format/width/height/stride/bo name (can this
be any EGLImage or just one from EGL_MESA_DRM_image?)
* turn a pixmap (from X?) into an EGLImage? (don't have the spec on
And then as of your changes we support:
* turn a GL 2D texture miplevel into an EGLImage
So I think this is the first time we need to deal with x,y offsets, and
I'd like us to be spec compatible. Tile alignment is really ugly -- not
only did original gen4 not support the fine-grained offsets, but even
gm45+ don't support sufficiently fine-grained alignments to point at
arbitrary miplevels. Depth buffers need to be aligned to 8x8, while
miptree levels only get aligned to up to the nearest 4 vertically.
If it's just going from a texture to a texture, I think we're safe on
gm45+ -- miptrees always have images aligned to at least 4x2. But for
renderbuffers I think we can weasel out by:
"If the GL is unable to create a renderbuffer using the specified
eglImageOES, the error INVALID_OPERATION is generated. If <image>
does not refer to a valid eglImageOES object, the error
INVALID_VALUE is generated."
and there's corresponding text for textures:
"If the GL is unable to specify a texture object using the supplied
eglImageOES <image> (if, for example, <image> refers to a multisampled
eglImageOES), the error INVALID_OPERATION is generated."
If we did that for these misalignment, I think I'd be satisfied with
dropping the actual rebasing. It would mean that everything but
miplevels of depth work on gm45+, and single-level textures work
pre-gm45. Faking shared storage for misalignment just seems
unreasonably hard unless we have a strong requirement for it.
Thinking about this all reminds me: We should error out for making
depth/stencil textures EGLimages as renderbuffers work. That way lies
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Size: 197 bytes
Desc: not available
More information about the mesa-dev